Unverified85% confidenceFactTime unknown

训练SWE-bench智能体的先进方法包括拒绝采样（Rejection Sampling）、DPO（Direct Preference Optimization）以及过程奖励模型（Process Reward Model）

1

Sources

85%

Confidence

Long-term

Relevance

6/1/2026

First Seen

Sources

SWE-bench开放评测环境与训练方案：AI编程智能体研发门槛大幅降低

twitterSWEbench

Related Entities

Rejection Samplingconcept DPOconcept Process Reward Modelconcept

Related Claims

VerifiedDPO（直接偏好优化）等对齐算法在降低训练成本的同时提升了模型在开放式指令下的表现一致性77% similar UnverifiedProcess Reward Model（过程奖励模型）对Agent的中间推理步骤给予细粒度反馈，缓解稀疏奖励导致的训练不稳定问题72% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/26949

API

curl https://kongchang.com/api/v1/knowledge/claims/26949

MCP

get_claim(id=26949)